Research On Image Fusion And Target Detection Algorithm Based On Multi-sensor Infrared Imaging System

Infrared imaging technology receives infrared radiation from targets and backgrounds for imaging.It has the characteristics of strong penetration,long detection distance,and all-weather detection.The development of infrared imaging technology makes up for the shortcomings of visible light imaging and has been widely used in remote sensing,night vision,thermal imaging,and hyperspectral imaging.With the expansion of the application field of infrared imaging technology,the amount of information provided by single-band infrared imaging can no longer meet the increasing demand for object detection capabilities.How to improve the quality of infrared images based on the existing infrared imaging hardware and software technologies has become the focus of research.Therefore,multi-band infrared image fusion has become a new trend in the development of infrared imaging technology.The fused infrared image enhances the understanding of the target and scene,and improves the accuracy of object detection.At present,most of the fusion algorithms of multi-band infrared images need to be closely combined with the design of the imaging optical system,and use the method of feature point detection for registration.This method limits the usage scenarios of the infrared imager,and cannot obtain good registration and fusion results in the case of large differences in image between different bands and poor image quality.Based on this,this paper focuses on the fusion of short-wave and long-wave infrared images,and focuses on the registration algorithm and the object detection technology of the fused image.This article first briefly summarizes the basic features of infrared images,the basic techniques of image registration and fusion,and the basic principles of object detection models.In order to address the problem of feature point deviation caused by inaccurate target edge positions in different bands,a registration algorithm was proposed to evaluate the similarity of the gray distribution of pixels between the images to be fused.Meanwhile,based on deep neural network,this paper uses the object detection model to further test the quality of the fused image.The innovative content and contributions of this paper are as follows:1)Analyze and establish a projection transformation model when cameras at different positions in the space shoot coplanar points.Based on this model,a registration algorithm based on search projection point registration and correlation coefficient evaluation is proposed,and several algorithm constraints are proposed according to practical applications to improve the calculation performance.The algorithm uses genetic algorithm to search for the best projection point instead of feature point matching to complete the image registration,avoiding interference caused by poor image quality.By calculating the correlation coefficients of the images in the registration area,the algorithm avoids the problem of feature point deviation caused by inaccurate target edges in different bands.At the same time,the proposed additional constraints greatly improve the performance of the algorithm.Experimental results show that the proposed algorithm has better and more stable performance in a variety of scenarios than the traditional feature point registration algorithm.This algorithm can be directly extended to the application of multiple cameras.2)Using the deep neural network model to process the fused image,it is verified that the registration algorithm proposed in this paper can effectively improve the performance of fused image object detection.Firstly,a large number of infrared images of two bands were collected and a fused image was generated to produce a single-band image and a fused image data set.Then,use the optimized feature extraction classification and region generation network to train the object detection model on the produced data set.Finally,cross-test different datasets on different models.It is verified that the model trained by the fused image achieves the highest average accuracy and intersection ratio on each data set,and indeed improves the object detection capability of a single band image.